Sensor device and method for calibrating a sensor device

Abstract

In a method for calibrating a sensor device, the sensor device generates corrected measurement data from measurement data of the sensor device using a sensor metamodel, wherein the sensor metamodel comprises a Gaussian kernel. Calibration measurement data is generated by the sensor device and / or an external sensor device under specified conditions. Calibration output data is generated by the sensor metamodel based on the calibration measurement data. The sensor device is configured by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adjusted using the calibration output data.

Description

[0001] The present invention relates to a sensor device and a method for calibrating a sensor device. State of the art

[0002] Sensor devices must be calibrated after production, as they always exhibit certain deviations from one another. This calibration can be performed based on a mathematical model.

[0003] Starting with a general model, the model's parameters are adapted to individual sensor devices during the calibration process. For this purpose, the sensor devices record specific measurement points, and the models are adjusted using calibration algorithms based on the measured values.

[0004] When the sensors are installed at the user's site, conditions change and the original calibration is no longer optimal. The user can then recalibrate the sensor device, but only to a limited extent due to the user's limited knowledge of the exact functioning of the sensor device and its calibration.

[0005] A method for conditioning Gaussian processes is known from Maddox et al., “Conditioning Sparse Variational Gaussian Processes for Online Decision-making”, 2021, arXiv:2110.15172. Disclosure of the invention

[0006] The invention provides a sensor device and a method for calibrating a sensor device having the features of the independent patent claims.

[0007] Preferred embodiments are the subject of the respective subclaims.

[0008] According to a first aspect, the invention accordingly relates to a method for calibrating a sensor device, wherein the sensor device generates corrected measurement data from measurement data of the sensor device using a sensor metamodel, and wherein the sensor metamodel comprises a Gaussian kernel. Calibration measurement data are generated by the sensor device and / or an external sensor device under predetermined conditions. Calibration output data are generated by the sensor metamodel based on the calibration measurement data. The sensor device is configured by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adjusted using the calibration output data.

[0009] According to a second aspect, the invention accordingly relates to a sensor device comprising a computing device that generates corrected measurement data from measurement data of the sensor device using a sensor metamodel, wherein the sensor metamodel comprises a Gaussian kernel. The computing device generates calibration output data through the sensor metamodel based on calibration measurement data. The calibration measurement data are generated by the sensor device and / or an external sensor device under predetermined conditions. The computing device calibrates the sensor device by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adjusted using the calibration output data. Advantages of the invention

[0010] With sensor devices, users usually have to adjust and recalibrate them for specific applications. This involves effort and expense. If this is not done, the sensors' performance is usually impaired. The invention enables automatic calibration of the sensor device, which can improve the accuracy of the measurements.

[0011] Using the Gaussian kernel, the sensor metamodel can learn a probability function that represents a distribution based on the mean vector (i.e., average) and the covariance matrix. The sensor device is calibrated by initially training the sensor metamodel with the Gaussian kernel for many different sensor devices, resulting in a general function that best represents the entire data domain. This general function corresponds to the sensor metamodel, which is initially suitable for the multitude of sensor devices of a certain type, which is expressed by the term "meta."

[0012] The sensor metamodel is then adapted, meaning the learned function is conditional. This means that the mean vector and the covariance matrix of the Gaussian kernel are adjusted using the calibration output data, typically reducing the covariance. The sensor metamodel is thus conditioned on a specific sensor device. This makes it possible to optimize the sensor metamodel for the specific sensor device.

[0013] The method requires little computing power and can be implemented with high accuracy. The advantage of this low computing power is that only matrix multiplications and additions are required during the calculation. This is also possible on relatively small processors, unlike, for example, a gradient step.

[0014] Using adaptive methods, the sensor metamodel is adapted to new conditions. For example, a sensor metamodel can be generated that is universally suitable for and covers several different tasks. Using different methods, this sensor metamodel can then be adapted to specific individual tasks (i.e., a specific sensor device).

[0015] According to a further development of the method for calibrating the sensor device, the sensor metamodel is trained using training measurement data before generating the calibration sensor data, wherein the training measurement data is generated by a plurality of training sensor devices. The training sensor devices are preferably of the same type as the sensor device to be calibrated.

[0016] According to a further development of the method for calibrating the sensor device, the sensor metamodel comprises an artificial neural network. The sensor metamodel can thus be trained to be usable for many sensor devices with as little training measurement data as possible.

[0017] According to a further development of the method, the Gaussian kernel is arranged in an output layer of the artificial neural network.

[0018] According to a further development of the method for calibrating the sensor device, the Gaussian kernel is either an exact Gaussian kernel or a sparse variational Gaussian process kernel. Gaussian processes determine an exact kernel, which is intended to learn the distribution precisely during training, while sparse variational Gaussian processes approximate this distribution.

[0019] An exact Gaussian kernel can be advantageous for a small amount of data, provided it sufficiently covers the test data domain, and a sparse variational Gaussian process kernel for a large number of data points, as it requires fewer mathematical operations and therefore less computing power. The condition, i.e. the adaptation of the sensor metamodel, can be calculated differently for the two methods. With the sparse variational Gaussian process kernel, conditional coefficients can be used to calculate the mean vector and the covariance matrix of the kernel. These coefficients are trained using the data during the training phase. A condition can then be implemented by calculating these coefficients, using new input data and the previous covariance matrix, i.e. the old conditional coefficients.

[0020] A sparse variational Gaussian process kernel has the advantage that it does not learn the exact distribution, but merely approximates a distribution. Since the training dataset typically cannot represent the entire possible data space, this is advantageous because it typically approximates a larger data space than that of the training dataset. If a condition is later applied, the data space can also be specified for sensors that are not represented in the data space.

[0021] According to a further development of the method for calibrating the sensor device, the sensor metamodel receives the measurement data of the sensor device as input data and outputs a measurement error of the measurement data as output data. The sensor device can use the output data to generate and output corrected measurement data.

[0022] According to a further development of the method for calibrating the sensor device, the sensor metamodel receives the measurement data of the sensor device as input data and outputs the corrected measurement data as output data.

[0023] According to a further development of the method for calibrating the sensor device, the sensor device is calibrated following a manufacturing process for the sensor device. Calibration of the sensor device can occur, for example, after the sensor device has been soldered onto a circuit board. During soldering, the sensor device experiences external stress, which leads to a change in the behavior of the sensor device. With the help of adaptive adjustment of the sensor metamodel, it can be adapted to the changes, and the sensor device demonstrates improved performance.

[0024] According to a further development of the method for calibrating the sensor device, the sensor device is calibrated multiple times. This allows lifetime effects to be compensated for.

[0025] According to a further development of the method for calibrating the sensor device, the calibration measurement data is generated during normal operation. This eliminates the need for special test signals as data. For example, a yaw rate sensor can be calibrated by generating measurement data without an applied yaw rate. The measured yaw rate must then disappear so that possible deviations are known.

[0026] According to a further development of the method for calibrating the sensor device, the sensor device is a microelectromechanical (MEMS) inertial sensor for measuring angular rates or accelerations, a pressure sensor, or a humidity sensor. These are often soldered onto different printed circuit boards and are used in various situations where, for example, humidity, bending stress, or temperature effects impair performance. Using the adapted sensor metamodel, which adapts to these situations, increases performance.

[0027] Further advantages, features and details of the invention will become apparent from the following description, in which various embodiments are described in detail with reference to the drawings. Short description of the drawings

[0028] They show: Fig. 1 is a schematic block diagram of a sensor device according to an embodiment of the invention; and Fig. 2 a flowchart of a method for calibrating a sensor device according to an embodiment of the invention.

[0029] The numbering of procedural steps is for clarity and generally does not imply a specific chronological order. In particular, several procedural steps can be performed simultaneously. Description of the embodiments

[0030] Fig. 1 shows a schematic block diagram of a sensor device 1. The sensor device 1 comprises a sensor component 3 and a computing device 2. The computing device 2 can have at least one processor, a microprocessor, an integrated circuit and / or an application-specific integrated circuit, as well as at least one memory.

[0031] The sensor component 3 generates measurement data. The invention is not limited to a specific type of sensor device 1. For example, the sensor component 3 can be a temperature sensor, a pressure sensor, an acceleration sensor, a humidity sensor, or the like.

[0032] In addition, further components 5 can be provided, which provide additional data to the data of the sensor component 3. For example, the further components can include an additional temperature sensor or other sensors that do not detect the target signal of the sensor device 1, but correlate with external stress influences.

[0033] A sensor metamodel is stored in the at least one memory of the computing device 2, which receives the measurement data from the sensor component 3 as input data. Based on the input data, the sensor metamodel generates output data that can be used to correct the measurement data.

[0034] The output data may, for example, include an (estimated) measurement error of the measured data as the initial data. The output data may also include the measured data already corrected for the (estimated) measurement error.

[0035] Using the sensor metamodel, the computing device 2 thus generates corrected measurement data from the measurement data of the sensor device 1. These corrected measurement data can then be output for further processing or evaluation.

[0036] The sensor metamodel comprises a Gaussian kernel. According to one embodiment, the sensor metamodel comprises an artificial neural network. The Gaussian kernel can be located in an output layer of the artificial neural network. The Gaussian kernel can be an exact Gaussian kernel or a sparse variational Gaussian process kernel.

[0037] The computing device 2 generates calibration output data through the sensor metamodel based on calibration measurement data. The calibration measurement data is generated by the sensor device 1 and / or an external sensor device 4 under predetermined conditions. For example, the sensor component 3 of the sensor device 1 can generate calibration measurement data under known conditions. Since the conditions are known, the calibration measurement data can be labeled, i.e., the actual, expected measurement data can be determined.

[0038] If the sensor device 1 is, for example, an acceleration sensor, the sensor device 1 can be placed on a flat surface. The acceleration in the z-direction must then correspond to the acceleration due to gravity. The expected measurement data (in this case, an acceleration in the z-direction equal to the acceleration due to gravity) can be assigned to the measurement data determined by the sensor component 3 (i.e., the calibration measurement data). Calibration output data can be generated based on the expected measurement data. This corresponds to the output data that the sensor metamodel determines based on the calibration measurement data provided as input data.

[0039] If, for example, the sensor metamodel outputs the measurement error as output data (corresponding to the calibration output data),

[0040] Furthermore, ground truth data is generated, corresponding to the difference between the expected measurement data and the calibration measurement data, i.e., the actual measurement error of the sensor component 3. This ground truth data is compared with the calibration output data to train the sensor metamodel.

[0041] The computing device 2 calibrates the sensor device 1 by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adjusted using the calibration output data. The sensor metamodel can be trained, for example, using labeled data. The calibration measurement data can be the input data, for example, and the calibration output data can be the output data calculated from it using the sensor metamodel. These are compared with labels, i.e., ground truth data.

[0042] When training or adapting the sensor metamodel, it can be specified that only the Gaussian kernel is trained, which can be achieved through conditional training. This changes the mean vector and covariance matrix of the Gaussian kernel. For example, the mean vector can be shifted during adaptation. The covariance matrix is ​​typically changed in such a way that the covariance is reduced.

[0043] The sensor metamodel can first be trained using training measurement data before generating the calibration sensor data. The training measurement data can be generated, for example, by a plurality of training sensor devices 1.

[0044] This will be explained in more detail in the procedure described below.

[0045] Fig. 2 shows a flowchart of a method for calibrating a sensor device, in particular the one shown in Fig.1 and described above. Conversely, the sensor device 1 described above can be configured to be calibrated using the method described below. First, a sensor metamodel is to be generated for this purpose.

[0046] In a first step S1, training measurement data is generated for this purpose. These data can be generated by training sensor devices 1 even before series production of the sensor device 1. The training sensor devices 1 can have the same structure as the sensor device 1, i.e., be of the same sensor type. The training sensor devices 1 can first be calibrated.

[0047] During the development of the sensor devices 1, measurements are taken under different conditions. These conditions are intended to simulate the potential application cases, but do not have to be identical. Examples include sensor devices 1 in application areas with high humidity, high temperature, or a combination of different conditions. The measured data is divided into baseline test data and calibration data. The baseline test data D B are measurement data for stress effects in the different conditions, while the calibration data D C Measurement data is what is available at the beginning of a typical use case.

[0048] For example, D B Data of the soldered sensor device 1 at high humidity, while for D CData from the soldered sensor device 1 at low humidity are included. Both data sets contain both the measured sensor-internal signals x and the ground-truth values ​​y. The sensor-internal signals are signals that are available to the sensor device 1 at any time during an application.

[0049] The internal sensor signal can, for example, be a quadrature in a MEMS gyroscope. This signal is generated by coupling the driven movement of a gyroscope mass of the gyroscope to the detection movement and is caused by mechanical inaccuracies in the manufacturing process. Since the signal depends on the mechanical design, as well as the sensitivity and sensor error of the gyroscope, it changes under external stress, for example, and can be used to correct sensor errors.

[0050] Changing the quadrature can be used to calibrate the sensor without knowing the exact error. This can be used to improve, for example, the offset, sensitivity, cross-sensitivity, vibration robustness, noise, and bias instability of the sensor device 1. Furthermore, in addition to quadrature, other signals such as those from dedicated stress sensors or temperature sensors can be used as monitor sensors, or other internal sensor signals such as frequencies, quality factors, phases, and the like.

[0051] In one embodiment, monitor diodes are provided in laser modules for optical path length measurement sensors. The monitor diode is explicitly incorporated into the design of the sensor device to measure fluctuations in laser intensity. A measurement signal from the monitor diodes can be used to recalibrate the sensor without knowing the exact sensor error.

[0052] In this embodiment, in addition to the use of a monitor diode, an operating current of the laser module can also be used, which can be read out, for example, via a shunt resistor, in order to improve the number of signals and thus the accuracy of the algorithm.

[0053] In step S2, a sensor metamodel is trained. The sensor metamodel includes a Gaussian kernel. The Gaussian kernel can be an exact Gaussian kernel or a sparse variational Gaussian process kernel.

[0054] A Bayesian algorithm can be used, in which the Gaussian kernel learns a distribution function and can be conditioned to adapt to a specific task. Here, the individual sensor device 1 is defined as a task.

[0055] Using the training, the sensor metamodel can then be found for a sensor type of sensor device 1 and used as a compensation model. This compensation model estimates the sensor error.

[0056] The sensor metamodel can comprise a neural network or a differentiable mathematical function, with a Gaussian kernel as the terminator. For example, the Gaussian kernel can be located in an output layer of the neural network.

[0057] The sensor metamodel can be trained based on the training measurement data using a classical training algorithm. The sensor metamodel can be assigned to a function f θ with the parameters θ and can calculate the sensor error based on sensor-specific measurements or signals and mathematical models, ie fθ(x)=y, where x denotes the input data (e.g. the measurement data of the sensor device 1) and y denotes the output data (e.g. the measurement error or the corrected measurement data).

[0058] The sensor metamodel corresponds to a general model for all sensor devices 1 of a specific sensor type. Training may include an optimization step, which begins with a forward pass, i.e., sending the data through the sensor metamodel to obtain the output data.

[0059] The performance is then determined using, for example, a loss function such as the exact marginal log-likelihood of the sensor error L DB∪DC : L=pfθ(y|x)=∫p(y|fθ(x))p(fθ(x)|x)dfθ where p denotes the distribution function. Finally, a training process can be performed using a gradient step, and the parameters of the meta-model can be changed, for example, according to: θ=θ−α∇θLDB∪DC(fθ)

[0060] Here, α denotes a scaling parameter and corresponds to the learning rate. This step is repeated for each training sensor device.

[0061] In a further embodiment, the sensor metamodel is conditioned before the training step, based on the calibration data D C .

[0062] Series production of the sensor device 1 then begins. A specific sensor device 1 can be calibrated, for example, after final installation. The sensor device 1 is designed to generate corrected measurement data from the sensor device 1 using the trained sensor metamodel. The metamodel is thus implemented on the sensor device 1, for example, on a microcontroller of the sensor device 1.

[0063] In a step S4, calibration measurement data are generated by the sensor device 1 and / or an external sensor device 4 under predetermined conditions.

[0064] In step S5, the sensor metamodel generates calibration output data based on the calibration measurement data by receiving the calibration measurement data as input data. The calibration output data are the corresponding output data.

[0065] Ground truth data is then determined based on knowledge of the conditions.

[0066] In step S6, the sensor device 1 is calibrated by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adjusted using the calibration output data and the ground truth data. The sensor metamodel is thereby conditioned.

[0067] During adaptation, the model can first be conditional using the calibration measurement data and calibration output data. To do this, the conditional coefficients can be recalculated, thus recalculating the mean vector and the covariance matrix of the Gaussian kernel, as described in Maddox et al., "Conditioning Sparse Variational Gaussian Processes for Online Decision-making," 2021, arXiv:2110.15172.

[0068] Thus, all future predictions of the sensor metamodel are conditioned by the calibration data and the probability for a prediction y for certain input values ​​x changes to p(y|x,Dc,θ).

[0069] The adapted or optimized sensor metamodel achieves significantly higher performance for the specific application situations.

[0070] The condition can be repeated after some time to compensate for lifetime effects. Ground truth values ​​can be generated again for this purpose. For MEMS gyroscopes, for example, this is possible by leaving the gyroscope at rest, so that it is known that no angular rate is applied.

[0071] The sensor device 1 is thus adapted to a specific application situation. This can increase the performance of the individual sensor device 1 for the respective application. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] Maddox u.a., „Conditioning Sparse Variational Gaussian Processes for Online Decision-making“, 2021, arXiv:2110.15172 [0005, 0067]

Claims

[1] Method for calibrating a sensor device (1), wherein the sensor device (1) generates corrected measurement data from measurement data of the sensor device (1) using a sensor metamodel, and wherein the sensor metamodel comprises a Gaussian kernel, comprising the steps: Generating calibration measurement data by the sensor device (1) and / or an external sensor device (4) under predetermined conditions; Generating calibration output data by the sensor metamodel based on the calibration measurement data; and Calibrating the sensor device (1) by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adapted using the calibration output data. [2] The method according to claim 1, wherein the sensor metamodel is trained using training measurement data prior to generating the calibration sensor data, the training measurement data being generated by a plurality of training sensor devices (1). [3] The method of claim 1 or 2, wherein the sensor metamodel comprises an artificial neural network. [4] The method of claim 3, wherein the Gaussian kernel is arranged in an output layer of the artificial neural network. [5] Method according to one of the preceding claims, wherein the Gaussian kernel is an exact Gaussian kernel or a sparse variational Gaussian process kernel. [6] Method according to one of the preceding claims, wherein the sensor metamodel receives the measurement data of the sensor device (1) as input data and outputs a measurement error of the measurement data as output data. [7] Method according to one of the preceding claims, wherein the sensor metamodel receives the measurement data of the sensor device (1) as input data and outputs the corrected measurement data as output data. [8] Method according to one of the preceding claims, wherein the sensor device (1) is calibrated following a manufacturing process of the sensor device (1). [9] Method according to one of the preceding claims, wherein the sensor device (1) is calibrated several times. [10] Sensor device (1), comprising: a computing device (2) which is designed to generate corrected measurement data from measurement data of the sensor device (1) using a sensor metamodel, wherein the sensor metamodel comprises a Gaussian kernel; wherein the computing device (2) is designed to generate calibration output data by the sensor metamodel based on calibration measurement data, wherein the calibration measurement data are generated by the sensor device (1) and / or an external sensor device (4) and / or a sensor component (5) of the sensor device (1) under predetermined conditions; and wherein the computing device (2) is further configured to calibrate the sensor device (1) by adapting the sensor metamodel, wherein a mean vector and a covariance matrix of the Gaussian kernel are adapted using the calibration output data.

Images